An etiologic prediction model incorporating biomarkers to predict the bladder cancer risk associated with occupational exposure to aromatic amines: a pilot study

The present study includes the “cases” arm stemming from of an earlier hospital-based case-control study fully described in previous publications [17‐20]. Inclusion criteria were being male, aged 20–80, Italian. Cases were all 199 newly diagnosed, histologically confirmed BC patients, admitted to the Urology Departments of two large hospitals from 1997 to 2000. Controls were all 214 non-neoplastic urological patients matched to cases by age (±5 years), period and hospital of admission. A written informed consent was obtained from each subject; the local Ethical Committee approved the study.

Peripheral blood lymphocytes (PBLs) were collected and automated DNA extraction was performed according to Extragen kit (Extragen BC by TALENT) [17]. Genotyping of glutathione S-transferase M1 (GSTM1) null, GSTT1 null, GSTP1 I105V, N-acetyltransferase 1 (NAT1) fast, NAT2 slow, cytochrome P450 1B1 (CYP1B1) V432 L, sulfotransferase 1A1 (SULT1A1) R213H, myeloperoxidase (MPO) G-463A, catechol-O-methyltransferase (COMT) V108 M, manganese superoxide dismutase (MnSOD) A-9 V, NAD(P)H:quinone oxidoreductase (NQO1) P187S, X-ray repair cross-complementing group 1 (XRCC1) R399Q, XRCC3 T241 M, and xeroderma pigmentosum complementation group (XPD) K751Q polymorphisms was assessed using Amplification Refractory Mutation System assay [17]. Bulky-DNA adducts were detected by 32P–postlabeling after Nuclease P1 enrichment and labelled adducts resolution on Thin Layer Chromatography (TLC) [20]. DNA adducts levels were measured as relative adduct level per 10⁸ nucleotides. A trained interviewer collected information on demographic variables, lifetime smoking history, coffee and other liquid consumption, dietary habits, lifetime occupation history by questionnaire. Job titles and individual activities, as well as occupational exposures to AAs, were blindly coded by an occupational physician according to methodology described in previous publication [21]. Occupations involving exposure to AAs were attributed to 11 International Standard Classification of Occupations (ISCO, International Labour Office, 1968) codes for job tasks (1–61.30: Painter, Artist; 3–70.20: Mail Sorting Clerk; 5–70.30: Barber-Hairdresser; 7–41.40: Mixing- and Blending-Machine Operator, Chemical and Related Processes; 8–01.10: Shoemaker, General; 8–11.20: Cabinetmaker; 8–73.70: Vehicle Sheet-Metal Worker; 9–01.35: Rubber Moulding-Press Operator; 9–31.20: Building Painter; 9–39.20: Brush-Painter, except Construction; 9–39.30: Spray-Painter, except Construction) and 11 International Standard Industrial Classification of all Economic Activities (ISIC, United Nations, 1968) codes for industrial activities (3240: Manufacture of footwear, except vulcanized or moulded rubber or plastic footwear; 3320: Manufacture of furniture and fixtures, except primarily of metal; 3521: Manufacture of paints, varnishes and lacquers; 3559: Manufacture of rubber products not elsewhere classified; 3819: Manufacture of fabricated metal products except machinery and equipment not elsewhere classified; 3824: Manufacture of special industrial machinery and equipment except metal- and wood-working machinery; 3843: Manufacture of motor vehicles; 5000: Construction; 9415: Authors, music composers and other independent artists not elsewhere classified; 9513: Repair of motor vehicles and motorcycles; 9591: Barber and beauty shops).

We investigated exposure to AA in all jobs held during lifetime, carefully assessing the level and the temporal aspects of such exposure according to standardized procedures [21]. Then, in order to achieve an optimal case definition [22], the critical values for time since first exposure (TSFE) and time since last exposure (TSLE) were chosen based on literature findings. In a cohort of Italian dyestuff workers [23], the risk of BC mortality decreased with increasing TSLE and became non-significant at ≥30 years since last exposure. Out of 19 BC patients observed in a Japanese dyestuff-plant, 17 showed a TSFE ≥20 years and 18 a TSLE ≤35 years [24]. Thus, the criteria for best-case definition were: TSFE higher than 20 years; TSLE lower than 35 years; length of exposure of at least 1 year; any value of cumulative exposure to AAs. The 15 BC cases complying with the above criteria were considered AA-related bladder cancer.

Smoking was a categorical variable with three levels: nonsmokers; former smokers from >20 years; current smokers and former smokers from less than 20 years. Life-long time-weighted average of cups/day of coffee and age at diagnosis were broken down in four classes according to the tertiles. DNA adducts were transformed in logarithm, and all values >1 were coded as 1 and otherwise as 0. Genetic biomarkers were coded as 0/1 variables as follows: GSTM1 (“NULL” variant =1, otherwise = 0); GSTP1 (“1A/1A” = 0, otherwise = 1); GSTT1 (“NULL” = 1); NAT1 (“S” = 1); NAT2 (“S” = 1); MPO (“A/A” = 1); COMT (“WW” = 1); MnSOD (“WW” = 1); NQO1 (“MM” = 1); CYP1B1 (“WW” = 0); XRCC1 (“G/G” = 0); XRCC3 (“C/C” = 0); XPD (“A/A” = 0). All the above variables became the predictors in a model of logistic regression in which the dependent variable was 1 for the 15 patients (cases) with AA-related BC (see above) and 0 for the other 184 BC patients (controls). A stepwise selection of independent variables was made using 0.10 as “p-to-enter” and 0.15 as “p-to-remove”. Therefore, from the same sample of 199 cases of BC, two algorithms were obtained (full model and short model) reporting for each regressor the OR with 95% CI and p-value. The best fitting model was chosen with measures of predictive power (R-square and area under the ROC curve) and GOF statistics (Pearson chi-square and Hosmer-Lemeshow test). The criterion was “the higher the better” for the former, and “p-value above 0.05” for the latter. The graphical outputs of the ROC curves were obtained and the AUC was interpreted according to the classification proposed: 0.5 (not informative test); 0.5–0.70 (inaccurate test); 0.7–0.9 (moderately accurate test); 0.9–1.0 (highly accurate test); 1 (perfect diagnostic test) [25]. A statistical test comparing the equality of AUCs was also calculated. Lastly, using the prediction equation we obtained a new variable containing the model-predicted probability of a positive outcome; the same computer program provided the “best cut-off” of predicted probability [26] that maximized the difference between BC patients with or without AA-related disease in both models. Using such value we built the classification table (true positive, false positive, true negative and false negative) from which we calculated sensitivity, specificity, positive and negative predictive values and diagnostic accuracy of each model. Purely statistical measures for comparing two risk prediction models have, however, limited use for medical decision making because they do not incorporate harms and benefits related to treatment decisions arising from the risk prediction model [27]. To evaluate whether clinical use of prediction models, diagnostic tests, and molecular markers would do more good than harm, a simple type of decision analysis (net benefit, NB, approach) has been used [28]. The NB depends on the benefit B, the cost C, the prevalence P of the outcome, and the risk threshold, R, which expresses the model’s classification accuracy, that is the ability of the risk model to assign high risks to cases and low risks to controls. The key quantity is R, which is a function of the harms and benefits of the possible outcomes without detailed specification of harms and benefits [29]. Therefore, the NB to the population of using the risk model is:

where:

TPR_R = true positive rate, also called sensitivity;

FPR_R = false positive rate, also called one minus specificity;

P = probability of diseases at a given time, also called prevalence;

R = risk threshold, also called model-predicted probability of a positive outcome.

Net benefit can be plotted against a range of R, in what is called a “decision curve”. Decision curves are now widely used in the literature [28]. In the present study, however, wider effects on NB were observed with variations of P rather than of R.

In the present one-sample study, the sample size was estimated based on Fisher’s z test assuming a correlation 0.28 (that between DNA adducts and the 0/1 variable “presence/absence of occupational AA-related BC”) and a significance level of 0.05. For a two-sided hypothesis test, the estimated sample size was 98 or 130 patients setting the power at 0.80 or 0.90, respectively [30]. The actual number of BC patients was 199. All statistical analyses were performed with STATA 13.